Global Poverty Estimates: Present and Future

Global poverty estimates: Present and future

1 Shatakshee Dhongde

1 Rochester Institute of Technology, Rochester, NY, USA 2* Camelia Minoiu [email protected]

2 IMF Institute, Washington, DC, USA * Corresponding author October 2010 [email protected]

BWPI Working Paper 133

Brooks World Poverty Institute

ISBN : 978-1-907247-32-3

Creating and sharing knowledge to help end poverty

www.manchester.ac.uk/bwpi

Abstract

We review the recent empirical literature on global poverty, focusing on key methodological aspects. These include the choice of welfare indicator, poverty line and purchasing power parity exchange rates, equivalence scales, data sources, and estimation methods. We also discuss the importance of the intra-household resource allocation process in determining within-household inequalities and potentially influencing poverty estimates. Based on a sensitivity analysis of global poverty estimates to different methodological approaches, we show that existing figures vary markedly with the choice of data source for mean income or consumption used to scale relative distributions; and with the statistical method used to estimate income distributions from tabulated data

Keywords: global poverty, household surveys, national accounts, tabulated data

Shatakshee Dhongde is at the Department of Economics, Rochester Institute of Technology, Rochester, NY, USA.

Camelia Minoiu is at the IMF Institute, International Monetary Fund, Washington, DC, USA.

Acknowledgements The authors are grateful to participants at the CPRC International Conference 2010, ‘Ten years of war against poverty: What have we learned?’ held at the Brooks World Poverty Institute, University of Manchester (8-10 September) for useful discussions and comments. The views expressed in this Working Paper are those of the authors and do not necessarily represent those of the IMF or IMF policy.

Introduction Global poverty monitoring has been brought to the forefront of the international policy arena with the introduction of the Millennium Development Goals (MDGs). The first MDG proposes the elimination of severe poverty globally, and has been formulated as the goal of ‘halving the proportion of people with an income level below $1/day between 1990 and 2015’. The adoption of the MDGs has led to renewed interest in estimating poverty at the national, regional and global level. Nevertheless, the estimation of poverty at the global level is made difficult by data limitations and methodological challenges. For example, poverty estimates tend to be sensitive to the consumer price indices used to update poverty lines; the purchasing power parity (PPP) exchange rates required to make incomes and expenditure levels comparable across countries; and the statistical techniques employed to estimate income distributions.

Existing global poverty assessments put forth conflicting conclusions about the extent of poverty and the pace of poverty reduction. In this paper, we review the recent empirical literature on global poverty, focusing on key methodological issues. To analyse the sources of the discrepancies in estimates across different studies, we undertake a sensitivity analysis of global poverty rates and counts to changes in methodological approach. Specifically, we use a dataset with distributional information for 65 developing countries to estimate poverty in 1995 and 2005, alternating either the estimate of mean income that anchors national relative distributions, or the statistical technique used to estimate the income distribution from tabulations.

Our results suggests that a large share of the variation in estimated poverty levels and trends is attributable to the choice of surveys or national accounts as the primary data source for mean income (consumption). The estimation method for the income distribution from tabulations is also quantitatively important, although the trend of falling poverty over the past decade appears to be robust to the methodological approach. These findings suggest that efforts to uniformise data collection practices across countries, and to compile individual records from surveys into large- scale databases, are essential steps for advancing the debate and improving future global poverty statistics.

Some of the issues pertinent to our assessment have been discussed in the literature. For instance, Anand and Segal (2008) review the literature on global income inequality, highlighting the reasons that undermine confidence in existing estimates: measurement error in national accounts, survey data, and within-country price data; index numbers and multilateral comparisons for PPP estimates; and the lack of comparability of survey data across countries. Since poverty and inequality are different ways of analysing the income distribution, most of the issues they discuss are also relevant to the measurement of global poverty. Our contribution is to bring to the fore aspects that so far have remained understudied, such as prevailing data sources for estimates of mean income or consumption (e.g., surveys or national accounts) and the type of available data (tabulations or individual records).

The rest of the paper is structured as follows. In Section II we discuss the steps involved in assessing poverty globally, focusing on the task of conceptualising poverty and operationalising the definition. Section III describes key empirical and methodological choices in global poverty 3 assessments. Section IV presents a history of global poverty studies and discusses the existing estimates. In Section V we assess the sensitivity global poverty estimates to methodological choices, and we conclude in Section VI.

I. Conceptualising poverty

Global poverty assessments involve the following steps. The first step is to define poverty, i.e., to choose criteria based on which the poor can be identified. Sen (1976, 1993) defined poverty as absolute deprivation in terms of individual capabilities – the potential or personal advantage that an individual can attain. Accordingly, an individual’s standard of living is reflected by his capabilities rather than the number of commodities she possesses or the level of utility she derives from consuming those commodities. The capability approach, though conceptually appealing, has proved difficult to apply empirically.1

Data: Objective vs. subjective

Most of the empirical studies on global poverty take the objective approach to assessing wellbeing, using quantitative variables such as income, consumption or indicators of child nutrition and health. Data on self-reported assessments regarding living conditions remain extremely scarce at the global level, with notable advances made by the 2000 World Development Report (WDR) ’Voices of the Poor’, which described the views on poverty of 60,000 individuals (World Bank, 2000), and Deaton’s (2008) study of self-reported life satisfaction in 120 countries based on Gallup polls.2 Deaton (2010) argues that subjective data from the Gallup surveys provide a potentially useful indication of human wellbeing, and could more prominently be used to assess global trends.

Poverty concept: Absolute vs. relative

In global analyses, the focus has traditionally been on absolute poverty, which concerns the cost of meeting a given standard of living (Bhalla, 2002; Chen, Datt and Ravallion, 1994; Chen and Ravallion 1997, 2001, 2004, 2007, 2008; Kakwani and Son, 2006; Sala-i-Martin, 2006; Pinkovskiy and Sala-i-Martin, 2009). In contrast, relative measures of poverty aim to capture an individual’s inability to participate in society, and are defined in relation to the overall distribution of income. Thus, the (relative) poverty line is anchored to the mean (median) income level. Relative poverty assessments are more common in richer countries. For example, the official poverty line for countries in the European Union is set at 60 percent of median income (Trinczek, 2007). Global relative poverty has empirically been studied by Ravallion and Chen (2009) and Nielsen (2009).

Poverty concept: Uni-dimensional vs. multi-dimensional

Existing global poverty studies typically focus on uni-dimensional poverty, which is assessed based on a single welfare indicator (such as income or consumption). Multi-dimensional poverty is measured through a more inclusive approach that captures multiple dimensions of wellbeing (e.g.,

1 See Reddy, Visaria and Asali (2009) for an illustration. 2 See also Gasparini and Gluzmann (2009) for a recent poverty assessment for Latin America and the Caribbean using Gallup World Poll data for 2006. 4 access to basic social services, educational attainment, health status, availability of shelter and participation in the labour market). However, micro-data of the kind required to estimate multi- dimensional poverty across countries remain sparse.3 An early attempt to estimate global poverty in a multi-dimensional framework (including income, health and education) was the Human Development Index (HDI) – first published for a large number of countries in the 1997 Human Development Report (HDR). The most recent HDR proposes a Multi-dimensional Poverty Index that draws on a larger number of living standard, health and education variables to assess deprivation at the household level in 104 countries (Alkire and Santos, 2010).

Welfare indicator: Income vs. consumption

In the traditional money-metric approach to global poverty measurement, poverty is assessed based on income or consumption. Choosing between the two measures of welfare often depends on the availability and quality of existing data. Countries follow different practices when conducting household surveys. For example, Latin American and Central and East European countries are more likely to collect data on income, whereas surveys in Asian, African and Middle Eastern countries focus on consumption (Chen and Ravallion, 2004). Global poverty estimates are derived by aggregating national income and expenditure data after making adjustments to either of the two variables.4 Nevertheless, both income and consumption variables suffer from substantial measurement error (Deaton, 2003), as discussed further in Section III.

Poverty lines: National vs. international

If national poverty lines were consistent with a uniform definition of poverty, the number of global poor could be estimated by adding up corresponding poverty headcounts. In practice, this is not feasible because national poverty lines often correspond to different definitions of poverty. Furthermore, only few countries monitor poverty systematically, leading to data gaps in national statistics.5 Global poverty is instead estimated using international poverty lines (such as $1/day and $2/day) that are translated into countries’ local currencies using PPP exchange rates, and that are moved backward and forward in time using national inflation rates. The $1/day poverty line6 is close to the average of PPP-adjusted national poverty lines of the poorest 15 nations in the world (Ravallion, Chen and Sangraula, 2008).7

Several studies have discussed the challenges posed by the international poverty lines and their conversion into local currencies, arguing that GDP PPPs are inadequate for poverty assessment as

3 Although not available for every country and every year, micro surveys that systematically collect non- income information across countries include the Demographic and Health Surveys and the Living Standards Measurement Surveys. 4 For example, Chen and Ravallion (2001) rescale mean income by one minus the saving rate to obtain mean consumption for countries that only undertake income surveys (Chen and Ravallion, 2001: 7). The rescaling, however, is found to have little impact on estimated trends or inter-regional comparisons. 5 For example, the World Development Indicators (WDI) 2009 database includes national poverty statistics for only five percent of the countries. 6 Refers to $1.00 at 1985 PPPs, $1.08 at 1993 PPPs, and $1.25/day at 2005 PPP. 7 In turn, these national poverty lines are typically anchored to a nutritional norm representing 2,100 calories per person per day plus a non-food allowance roughly equal to non-food expenditures of individual close to the caloric cut-off. 5 they do not reflect the consumption patterns of the poor8 and the poverty thresholds do not capture the real requirements of human beings (Reddy and Pogge, 2010). Ackland, Dowrick and Freyens (2008) have shown that the choice of price index to calculate the PPP exchange rates leads to large differences in global poverty counts. PPP estimates are also updated at long intervals through the International Comparison Program (ICP), which only surveyed prices of commodities and services world-wide in 1985, 1993 and 2005. (China participated for the first time in the 2005 ICP round.) The 2005 update led to substantial revisions of historical series of PPP-adjusted GDP for a number of countries. Furthermore, estimates of global poverty and inequality were revised upwards using the new PPPs (Chen and Ravallion, 2008; Milanovic, 2009),9 while world growth estimates were revised downwards over 2002–07 to reflect countries’ updated weights in global GDP (Elekdag and Lall, 2008).

To sum up, the standard approach to global poverty measurement focuses on uni-dimensional absolute poverty (where the welfare indicator aiming to capture the standard of living) and estimates it by applying international poverty thresholds to objective distributional data on income or consumption from household surveys.

II. Key methodological choices

Here, we focus on three methodological issues that are crucial for global poverty assessments, but which have received relatively less attention in the literature: (i) the use of equivalence scales to calculate individual income (consumption) from the household aggregate, and the presence of intra-household inequality in the allocation of resources; (ii) the choice of mean income estimate (from household survey vs. national accounts) that anchors national relative distributions; and (iii) the estimation method of the income distribution from tabulated data.

A. Equivalence scales and intra-household inequality

Although the relevant unit in global poverty assessments is the individual, data is often available only at the household level. To obtain consumption levels for each individual in a household, household consumption is typically divided by the number of household members. However, this procedure rests on the unrealistic assumption that all household members have equal consumption levels. In reality, the consumption requirements to achieve the same level of welfare of children and the elderly are different from those of adults.10 Equivalence scales, which compute the adult- equivalent consumption level of household members depending on their caloric intake, can be used to model consumption heterogeneity among household members. To obtain per capita consumption, total household consumption is then divided by the number of equivalent adults rather than the number of household members.

8 Deaton and Dupriez (2008) adjust PPP exchange rates by re-weighting the consumption basket to address this problem. 9 A number of recent studies have critically reviewed global poverty estimates based on the 2005 ICP PPPs (see, e.g., Himanshu, 2009; Klasen, 2009). 10 Furthermore, consumption of some goods can be economised among members, the household providing opportunities for economies of scale. Szekely et al. (2004) document the impact of assumptions regarding economies of scale in consumption on estimated regional poverty in Latin America. 6

There is no global poverty assessment to date that attempts to use consumption per equivalent adult, although national and regional analyses often do. One reason is that the necessary adjustments require access to individual records, whereas much data is only available in tabulated form (Shorrocks and Wan, 2008). A second reason is that there is no consistent source of equivalence scales across countries. Even if individual record data were available for the entire world, equivalence scales would have to be constructed for each country to account, e.g., for relative differences in the cost of children (Milanovic, 2005: 18-19).

While it is likely that using equivalence scales in the assessment of global poverty would change existing estimates, the magnitude of any adjustments is unclear ex ante. However, evidence from individual country studies suggests that the impact of incorporating equivalence scales in studies of global poverty could be important. For example, Buhmann et al. (1988) use data from the Luxembourg Income Study for 10 OECD countries to show that the use of equivalence scales, as well as the choice among them, systematically affects estimates of absolute and relative poverty. The degree of sensitivity varies across countries, although rank-orderings by poverty and inequality level are relatively robust (see also Haughton and Khandker, 2009; Burniaux et al., 1998). This leads the authors to conclude that ‘because of these sensitivities one must carefully consider summary statements and policy implications derived from cross-national comparisons of poverty and/or inequality’ (p. 140). The same conclusion is reached by Szekely et al. (2004), who assess the sensitivity of regional poverty estimates for Latin America to the choice of adult equivalence scales (among other parameters) and document a large range of variation in estimated headcount ratios depending on the mix of parameters employed.

In a global assessment of poverty, equivalence scales-related adjustments would be largest for countries that experienced the steepest changes in demographics and household characteristics (Burniaux et al., 1998; Betson, 1996, 2004). This is the case of China – the main driver of poverty reduction in the developing world.11 Between 1990 and 2005, the average household size in China fell from 3.4 to 3 (Chamon and Prasad, 2010: 98) and the share of the population older than 65 increased from 5.6 to 7.5 percent.12 We summarise the evolution of regional demographic trends in recent decades in Table 1. Overall, the share of young population has significantly fallen over the period, while that of population above 65 has markedly risen. East Asia and sub-Saharan Africa (SSA) have registered the largest and smallest demographic transformations, respectively. There are also important cross-regional differences in average household size, but little empirical evidence on how it changed over time.13

A related issue is that adult equivalence scales themselves are imperfect for estimating individual consumption, since the latter depends not only on demographics, but also on the nature of the resource allocation process within households. Adult equivalence scales assume that there is no household inequality. Although the empirical evidence on patterns of intra-household resource

11 Outside China, the number of poor has increased since 1981 for all poverty lines higher than $1.25/day (Chen and Ravallion, 2008). 12 WDI (2006). 13 A notable exception is Bongaart (2001), who analysed household structure in developing nations using 43 household surveys over 1990-1998. He concluded that convergence to smaller households was proceeding slowly in the developing world based on the available data. 7 allocation remains scant, the results of recent studies are informative. Lise and Seitz (2008) show that using standard equivalence scales (which ignore intra-household inequality) leads consumption inequality in the UK to be underestimated by 30 percent – a result explained by the fact that the gender earnings gap translates into a consumption gap inside the household. Furthermore, British data reveal that household inequality has declined markedly since the 1970s as the share of income earned by wives has increased. These findings suggest that true consumption distributions for women and men are markedly different from those obtained using adult equivalence scales.

These concerns are particularly relevant for developing countries where poverty rates are higher and the allocation of resources within households also depends on cultural norms. A study of the number of decision makers in Turkish households shows that while consumption patterns in the average Turkish household are consistent with a multi-person model, households in Eastern Turkey, where traditional values prevail, are ‘unitary’, in the sense that one decision-maker allocates resources to the other members of the household (Kapan, 2009). The single-decision- maker model cannot be rejected in subsamples of households in which women do not participate in the labour market or whose children are female. The genders of children and the wives’ outside options are therefore important determinants of the allocation of resources within the household. This has implications for the assessment of welfare in countries where labour market participation for women remains limited or where patriarchal values are dominant.

Could intra-household inequalities play an important role in the assessment of global poverty? Table 2 reports regional trends in gender inequalities in education and labour market participation since 1990. There are important regional differences in the male-to-female ratio of literacy and primary completion rates and not all regions are closing the education gap at the same rate. Substantial reductions in the gender education gap (measured by these indicators) were attained over 1990–2004 in the Middle East and North Africa (MENA) region, East and South Asia, and SSA. If education and employment opportunities play the role of shifters of female voice in the household, inequalities in the intra-household allocation of resources are also changing, and standard equivalence scales would fail to capture that. 14

In our sensitivity analysis we ignore the issues of equivalence scales and intra-household inequality due to insurmountable data limitations, recognising that they provide a potentially fruitful venue for future research as more and better survey data become available.

B. Survey vs. National accounts-based data

A second key methodological issue concerns the source of data for welfare indicators such as consumption: the estimates can be drawn either from household surveys (HS) or from national accounts statistics (NAS). HS are typically organised by national statistical agencies and collect information from representative households on consumption expenditures and/or personal disposable income. As a result, HS-based consumption can suffer from flaws in survey design

14 Another important determinant of intra-household inequality is the sex ratio. Recent papers document the marked worsening of the sex ratio imbalance in China over the past decades (Wei and Zhang, 2009), which could also play a role in determining the allocation of household resources and influence estimated poverty. 8

(Deaton and Grosh, 2000) and lack of representativeness, recall bias,15 underreporting among the poor, and poor response rates among the wealthy (Mistiaen and Ravallion, 2003; Deaton, 2005). Expenditure surveys are also more expensive to undertake than income surveys, as they require multiple visits to the participating household and consumption diary-keeping over a specified period. In contrast, NAS-based consumption is computed in the national accounts by subtracting net exports, investment, and government expenditure from national income.

How does the choice of estimate matter for poverty assessment? It has been documented that HS- based and NAS-based consumption differ both in level and growth rates. NAS consumption is higher than HS consumption (Deaton, 2001, 2005; Ferreira and Ravallion, 2008a) and grows faster than it. The level effect is due to the former including imputed rent on home-owners, imputed value of non-marketed items such as gifts, food produced and consumed at home, and consumption of non-profit organisations. In contrast, the latter can only imperfectly evaluate home-produced consumption, as market prices for self-produced goods can only be collected from distant (often urban) markets, and rents in rural areas with thin housing markets are imputed based on prices from vibrant (urban) rental markets. NAS consumption grows faster than HS consumption because it includes goods and services that are rarely consumed by the poor and because richer households are less likely to participate in surveys (Deaton, 2005).16 Furthermore, pure measurement error, differences in coverage, the presence of an informal sector, and differences in consumption deflators, cause further discrepancies (Ravallion, 2003).17

The World Bank measures consumption poverty and hence uses HS-based data whenever possible. Missing observations are interpolated across years using growth rates of per capita private consumption from the national accounts (Chen and Ravallion, 2008). Another approach is to scale HS estimates by the ratio between NAS income and HS income. Other studies circumvent this problem by relying solely on NAS data: Bhalla (2002) and Bourguignon and Morrisson (2002) employ NAS consumption as the welfare aggregate, whereas Sala-i-Martin (2006) uses NAS income. Deaton (2005) notes that if HS are wrong and NAS are correct, then using HS to estimate poverty will tend to underestimate the decline in poverty. Similarly, if HS are correct and poverty is estimated using NAS, then the results will overstate the pace of poverty reduction.

C. Estimation methods

The last methodological issue we discuss relates to the method of estimation of the income distribution. An important hurdle in estimating long-term trends in regional or global poverty is the lack of individual record data (unit data) for multiple countries and years. Efforts to undertake household surveys are often interrupted by conflict or undermined by poor statistical infrastructure. Surveys are sometimes available at periods far apart, and the individual records are unavailable in the public domain. Researchers often rely on published summary statistics – grouped frequency

15 For example, a substantial debate over the accuracy of Indian poverty estimates was caused by changes in the recall period (see, e.g., Deaton and Kozel, 2005 and Dhongde, 2007). 16 Ferreira and Ravallion (2008b) emphasise that the NAS-means method is unacceptable when doing an urban-rural poverty assessment. 17 Similar considerations arise when income poverty is estimated using per capita GDP from the NAS rather than per capita income from HS. 9 tables or ‘tabulated data’ – representing income (consumption) shares for a small number of population groups.18 Virtually all the studies on global poverty and inequality use a mix of individual records and tabulated data, especially for large countries such as China and India (Milanovic, 2005; Chen and Ravallion, 2008). Others derive national income distributions solely from summary statistics in (Bourguignon and Morrisson, 2002; Bhalla, 2002; Kakwani and Son, 2006; Sala-i- Martin, 2006; Ackland, Dowrick and Freyens, 2008; Pinkovskiy and Sala-i-Martin, 2009).

The global poverty literature proposes two methods for estimating national income distributions from tabulated data. The first approach is parametric: the Lorenz curve or the income density is parameterised using a simple functional form. For the Lorenz curve, the most widely used parameterisations are the General Quadratic (‘GQ’, proposed by Villasenor and Arnold, 1989) and the Beta (Kakwani, 1980).19,20 For the income density, the two-parameter log-normal distribution is the preferred candidate (Babones, 2003; Pinkovskiy and Sala-i-Martin, 2009), but other parameterisations have been proposed. These include a maximum entropy estimator for a density from the exponential family (Wu and Perloff, 2005, 2007) and the Generalised Beta-2 distribution parameterisation for the income density (Chotikapanich, Griffiths and Rao, 2007; Chotikapanich, Rao and Tang, 2007). Neither of the latter two approaches has been used in global poverty assessments.

The second approach for estimating national income distributions from tabulated data is non- parametric (Sala-i-Martin, 2006; Zhang and Wan, 2006; Ackland, Dowrick and Freyens, 2008). The method consists of applying the kernel density estimator on the tabulated data (expressed as mean incomes for several population groups), and has the advantage that no functional assumption needs to be made regarding the underlying data-generating process. Nevertheless, kernel density estimation requires specifying additional parameters (such as the kernel and bandwidth), which can have a large impact on the resulting estimate if applied to tabulated data rather than individual records.

Recent studies have assessed the performance of traditional methods in estimating the underlying distribution from tabulated data, and have concluded that parametric approaches provide more reliable estimates than non-parametric ones. For example, Minoiu and Reddy (2009) used Monte Carlo simulations on data from plausible income distributions to find that the GQ and Beta Lorenz curve parameterisations perform well in estimating poverty and inequality from tabulations, with

18 A comprehensive income distribution database (of income shares and Gini coefficients) is the UNU-WIDER World Income Inequality Database (WIID), which draws on multiple sources of information for developed, developing, and transition countries (Deininger and Squire, 1996; the Luxembourg Income Study; the Transmonee data by UNICEF/ICDC; Central Statistical Offices; and other research studies). Another source is the collection of survey-based data at the World Bank, which systematically compiles cross-country distributional data and disseminates national, regional, and global poverty statistics on its research studies and on its Povcalnet website (http://go.worldbank.org/NT2A1XUWP0). 19 Parametric approaches have been used in Bhalla (2002), Pritchett (2006), Kakwani and Son (2006), and the World Bank’s studies on global poverty. 20 Implementation of these methods can be done using the computational tools Povcal and SimSIP developed by the World Bank (Datt, 1998). Povcalnet poverty estimates are obtained using these Lorenz curve parameterisations. 10 biases rarely exceeding one percentage point.21 Further, Minoiu and Reddy (2008) empirically assessed the small-sample bias of the kernel density estimator to find that errors in the density and poverty estimates are often large, and depend crucially on the choice of bandwidth and on the position of the poverty line relative to the population median.

To our knowledge, no study to date has systematically assessed the variations in global poverty estimates to changes in the estimation method. An exception is Pinkovskyi and Sala-i-Martin (2009), who report correlation coefficients between national poverty rates corresponding to different parameterisations for the income density. While these correlations capture common trends in the series being compared, they preclude an assessment of level differences. We attempt to fill this gap in the sensitivity analysis presented in Section V.

III. Review of global poverty studies

A. Chronology

Global poverty assessments were undertaken for the first time at the World Bank, which started to systematically compile cross-country distributional data in the late 1970s.22 Earlier contributions include Paukert (1973), who tested the Kuznetz hypothesis using relatively comparable income distribution data for 56 nations. Two decades later, Ravallion, Datt and van de Walle (1991) estimated developing world poverty in 1985 using data from only 22 countries and an extrapolation model for 64 other nations. Chen, Datt and Ravallion (1994) and Ravallion and Chen (1997) expanded the coverage to 44 and 67 countries, respectively, to measure progress in reducing poverty between the mid-1980s and the early 1990s.

The first paper to rely entirely on survey data – a mix of individual records and tabulations – was Chen and Ravallion (2001), who assembled distributional information from 265 surveys in 83 developing nations.23 In their most recent study, Chen and Ravallion (2008) derived their poverty statistics from 675 nationally representative surveys in 115 developing nations. (For a chronology of global poverty studies since the late 1970s, see Table 3.) The scale of this study reflects the remarkable progress that has been made in compiling cross-country income distribution data over the past decades.

To illustrate how much information is available to study the long-run global distribution of income in public databases, we compiled data from the World Income Inequality Database (UNU-WIDER WIID 2008), Povcalnet, WDI, and the updated Dollar and Kraay (2002) dataset.24 In all, income

21 To further improve on these techniques, Shorrocks and Wan (2008) develop an algorithm which ensures that the simulated sample matches the moments of the tabulated data. 22 Distributional data for only 20 countries were published in the World Development Report between 1979 and 1995. The 1996 World Development Report included distributional data for 67 lower- and middle-income countries. 23 To fill in the gaps, countries without data were imputed their regional neighbours’ average poverty rate. 24 To construct the dataset, we started with the UNU-WIDER WIID dataset and retained all observations, regardless of data quality. Unique country-year observations from the other datasets are subsequently added to obtain an ‘augmented’ WIID world distribution dataset. Only a few observations available in the Dollar and Kraay (2002) and WDI are not already present in the other sources. The final dataset includes information on income shares for population quintiles and Gini coefficients (as reported in the original databases). 11 shares for population quintiles and Gini coefficients are available for 154 countries over 1960–2007 and are drawn from 3,031 underlying surveys (Table 4). While high-income countries have almost 30 surveys per country (out of 47 possible), middle- and low-income countries have 20 and 10 surveys on average. Survey coverage is best in South Asia and Latin America and worst in SSA and the MENA region. Data availability has increased markedly from some 30 country-surveys in the 1960s to 123 in the 1990s. The number of countries for which data is available rose from 17 in the 1960s to three times as many in the 1990s (Figure 1).

B. Global poverty estimates: What do they tell us?

Published poverty figures offer conflicting conclusions about the extent of poverty and the pace of poverty reduction.25 Take, for example, the latest two studies: Chen and Ravallion (2008) and Pinkovskiy and Sala-i-Martin (2009) (henceforth, ‘CR’ and ‘PS’). Both studies present estimates of the global income distribution, but differ markedly in terms of underlying data, interpolation techniques, choice of and data source for the welfare indicator, and estimation method. Table 5 summarises the authors’ choices. Key differences include the scope of the analysis (developing world vs. world) and the fact that CR estimates consumption poverty, while PS focus on income poverty (and adjust consumption shares accordingly to correspond to income shares). Once the income (consumption) shares are assembled, CR anchor the country-specific distributions mostly to HS consumption estimates, while PS use NAS per capita income (GDP). Finally, CR use a mix of individual records and tabulated data, on which they estimate the Lorenz curve using the GQ method, while PS rely solely on tabulations to estimate national income distributions using the log- normal parameterisation.

Both studies employ PPP estimates from the latest round of the ICP and estimate poverty relative to the standard thresholds $1/day, $2/day, etc. Chen and Ravallion (2008) estimate that in 2005 nearly 26 percent of the population in the developing countries was poor; the global poverty count fell by 520 million individuals since 1981. In contrast, Sala-i-Martin presents a $1/day poverty rate of 2.4 percent in 2005, corresponding to a reduction in the poverty count of almost 350 million individuals since 1981. Figure 2 plots estimates of the global poverty rate corresponding to different studies and PPPs. While level estimates vary substantially, most authors document a falling trend over the past decades, though the extent of the decline remains subject to debate.26

IV. Sensitivity analysis

Our approach is similar to Szekely et al. (2004), who investigate the sensitivity of regional poverty estimates in Latin America to a range of methodological choices (such as equivalence scales, economies of scale in consumption, and methods for treating missing/zero income and misreported entries). Our contribution is to focus on a different vector of methodological choices, which comprises the data source for estimates of mean income (or consumption) used to anchor the Lorenz curve; and the statistical technique used to estimate the income distribution from tabulated data.

25 For a comparative analysis of earlier global poverty assessments, see Quibria (2005). 26 See, e.g., Reddy and Minoiu (2007). 12

A. Survey- vs. national accounts-based poverty estimates

First we assess the sensitivity of global poverty estimates to changes in the data source of mean income (or consumption). The aim is to determine the variation in global poverty estimates attributable to the use of HS vs. NAS data while all the other parameters of the analysis are held constant. The difference this choice makes to the final estimates is not obvious from existing studies because, as discussed previously, these differ in many other dimensions.

Before proceeding to the results, we give a short description of the task of estimating poverty. When individual record data are unavailable, we must use tabulations that capture the country’s relative income distribution and come in the form of income shares for population quintiles or deciles. For example, we know the share of national income possessed by population quintiles, deciles, or any number of population groups. (Typically, income shares are available for at most 20 groups.) To obtain a distributional profile (i.e., the average income for each group) the relative distribution must be anchored to a mean level of income (or consumption) – which can be drawn either from HS or NAS. Since mean income is only used to scale the relative distribution, its effect is simply to shift the income distribution along the income axis (Figure 3).

Our main data source for is the Povcalnet database, which is particularly fit for our sensitivity analysis because it includes HS estimates of average consumption.27 In addition, Povcalnet includes relative distributions for a large number of countries.28 We take 1995 and 2005 as benchmark years, and select 65 countries for which distributional information is available in those years. 29 Poverty rates and counts are estimated for international poverty lines ranging between $1/day and $2.5/day; and the GQ parameterisation for the Lorenz curve. Note that we are largely replicating the World Bank methodology described in detail in Chen and Ravallion (2008).30 NAS income and consumption are taken from the Penn World Tables Mark 6.3 (Heston, Summers and Aten, 2009).

We report un-weighted and population-weighted summary statistics for PPP-adjusted consumption and income over 1995–2005 in Table 7. Mean consumption from surveys is the lowest estimate of the three aggregates, with NAS consumption exceeding it by a large factor. The gap between the two estimates has increased over time: while NAS consumption was larger than HS consumption by a factor of 1.6 or 1.9 in 1995 (for the weighted and un-weighted samples, respectively), this has increased to 1.8 or 2.3 by 2005. The level difference between survey consumption and NAS

27 Povcalnet reports either HS mean consumption or mean income, depending on the nature of the underlying survey. Mean consumption is available for two-thirds of the sample; when consumption is unavailable, we use income. (For a detailed discussion, see Chen and Ravallion, 2008: 15). 28 We find no systematic differences between the income and consumption shares (i.e., a regression of the shares against an indicator variable for the type of survey yields a statistically insignificant coefficient – the results are available upon request). Therefore, we use the raw data without further adjustments. 29 For countries with no information in 1995 or 2005, we use data from adjacent years, namely 1993–97 and 2003–07. For China, India and Indonesia, Povcalnet reports rural/urban distributional data, but there is no PPP-adjusted income (consumption). We therefore use national tabulations instead from alternative sources such as the WDI and WIID. Our sample thus covers slightly more than 70 percent of the total world population. (For the full list of countries, see Table 6.) 30 Nevertheless, the estimates we present should not be interpreted as unbiased estimates of the true extent of poverty, since we simply use a sample of countries for which both HS and NAS data are available to investigate robustness to changes in mean income (consumption) estimates. 13 income (GDP) is even higher. Furthermore, survey consumption has registered the lowest increase of the three aggregates, with an average annual growth rate of 0.9 percent over the period, compared to 3.1 percent for NAS consumption and 3.8 percent for GDP.31

Given that the mean income (consumption) estimates serve as anchor for the relative distributions of countries, these level- and growth differences between the various estimates will affect both the estimated extent and trend of global poverty over the period. The results confirm this intuition (Table 8). For example, in 2005, the $1/day consumption poverty rate was 29 percent using HS consumption, and only 5.9 percent using NAS consumption. The difference between HS- and NAS- based poverty estimates is large and increases still by 2005, reflecting the different growth rates of the various welfare aggregates. As a result, the pace of ($1/day consumption) poverty reduction varies between 16 percent and 32 percent, depending on the data source. While the falling trend of the poverty headcount ratio appears robust across poverty lines, poverty appears to have increased when we use the absolute headcount as the relevant measure. Specifically, HS-based consumption yields an upward trend in $2/day and $2.5/ day poverty, while NAS consumption and income yield falling poverty headcounts. The reason for this difference is that the growth rate of welfare measured from the surveys is insufficient to compensate for population growth and lead to a decline in poverty.32

We conclude that a large share of the variation in published estimates of poverty is attributable to the choice of data source for mean income (or consumption): HS vs. NAS. On average, survey consumption is (substantially) lower than its national accounts counterparts, generating (substantially) higher poverty rates. Furthermore, survey consumption (income) grows slower than NAS consumption (income), leading to a slower estimated pace of poverty reduction.

B. Estimation methods

The second choice regards the estimation method of income distributions from tabulated data. From the wide range of techniques available, we select six parametric and six non-parametric options, as follows: the GQ and Beta parameterisation for the Lorenz curve; the two-parameter log- normal and three-parameter Singh-Maddala parameterisations for the income density function; the Beta and log-normal parametric approaches with the Shorrocks and Wan (2008) correction; and the kernel density estimator with six alternative data-driven bandwidths. The parametric techniques are described in detail in Abdelkrim and Duclos (2007) and Datt (1998), while the non-parametric ones in Wand and Jones (1995). Before presenting the results, we briefly discuss each technique.

All of the methods above can be used to estimate the Lorenz curve or the income distribution from aggregate distributional information. The GQ and Beta approaches assume functional forms for the

31 These estimates are of comparable magnitudes to those reported by Deaton (2005), the differences being driven by a different sample composition and period of analysis; and different PPPs. Deaton uses 277 surveys over 1979–2000 to find that NAS consumption is on average 20 percent higher than survey consumption in the full sample that includes advanced economies; and grew over 1990–2000 at an average growth rate of 2.3 percent, compared to 4.5 percent for its NAS counterpart. Our analysis goes one step further, in that we examine the impact of these discrepancies on poverty estimates. 32 The results are consistent with the headcount estimates reported by Chen and Ravallion (2008) for higher poverty lines, which increased over the period 1981–2005. 14

Lorenz curve, the estimation of which involves a simple regression to estimate the parameters. Similarly, the log-normal and the Singh-Maddala functions are theoretical distributions that have been shown to approximate well real-world income data (Bandourian, McDonald and Turley, 2003). The procedure proposed by Shorrocks and Wan (2008) aims to improve on the initial distribution generated from the estimated Lorenz curve by correcting the initial sample of incomes to match the group means from the original data.

Finally, the kernel density estimator is a non-parametric smoothing technique aimed at estimating the income density. The simplest non-parametric technique for estimating the density is the histogram; the kernel density estimator produces a ‘smooth’ histogram. While the method has the advantage of not assuming a (potentially restrictive) functional form, it requires specifying a bandwidth which controls the smoothness of the estimated density. The bandwidth is the interval around each point of estimation where the estimator looks for information about the density at that point (income). The statistical literature proposes a number of ‘optimal’ bandwidths which minimise the mean squared error of the estimator, among which we select six examples. Importantly, different bandwidths lead to different results; hence it is good practice to assess the robustness of any non-parametric estimate to changes in the bandwidth.

We report global poverty rates and counts for 1995 and 2005 using three poverty lines ($1/day, $1.45/day and $2.50 day) and our original sample of 65 countries (Tables 9–10). The $1/day poverty headcount ratio varies in 1990 between 4.2 percent and 8.9 percent, or by a factor of 2.1. Similarly, in 2005 it varies by a factor of 3.2, depending on the method employed. The large variations in level-estimates of poverty across different methods are also apparent for the higher poverty lines. However, the falling trend in the poverty rate is robust across estimation methods, with declines ranging by 13 to 25 percentage points for the poverty lines considered. Similarly, the absolute headcount varies by between 200 and 237 million people in 1990; and by 131 to 272 million people in 2005. For this indicator, too, the trend of poverty reduction is robust across estimation methods, but the extent of the decline varies between 102 and 165 million people.

We conclude that the estimation methods of the income distribution from tabulated data play an important role in determining the estimated extent and trend of poverty. While the direction of the trend seems robust, the extent of poverty and the extent of the decline in poverty vary markedly with the method employed. This variation is informative, in that it can provide lower and upper bounds on poverty estimates, reflecting the uncertainty associated with the statistical technique used to generate them.

V. Conclusions

Global poverty monitoring is an important item on the international development agenda, with the first MDG aiming to reduce the share of ‘$1/day poor’ people in the world by 50 percent by 2015. In this paper, we reviewed the recent empirical literature on global poverty statistics, with the aim of constructing a coherent picture of the current state of knowledge and bringing to the fore issues that remain relatively understudied.

First, we characterised the standard approach to measuring poverty at the international level: this is uni-dimensional, it focuses on absolute poverty, and uses objective (income or expenditure) data. Attempts to depart from this approach – for example, to measure long-run multi-dimensional poverty or use subjective information – are rare and face important data-related obstacles. We also discussed key methodological issues in global poverty measurement, focusing on equivalence scale and the role played by intra-household inequalities for poverty and interpersonal inequality. Though conceptually important, and often acknowledged in existing global poverty assessments, these issues remain outside their scope because of data limitations.

We also undertook a sensitivity analysis of global poverty estimates to changes in the data source for mean income (or consumption) that is used to anchor relative income distributions; and to the statistical method used to estimate the income distribution from tabulated data that is often available in lieu of individual records. Our results suggest that a large share of the variation in estimated poverty levels and trends is attributable to the choice of surveys or national accounts as the primary data source for mean income (consumption). This choice alone appears to account for the bulk of level-differences in global poverty rates and counts across existing studies. The choice of statistical technique to estimate poverty from tabulated data is also important. Nevertheless, the falling trend of poverty over the past decades appears robust to the choice of methodological approach.

Overall, our results suggest that the debate concerning global poverty would benefit from improvements in data collection practices across countries, and the compilation of unit-record data from surveys into large-scale databases. Promising areas of future research include assessing long-term trends in absolute and relative human wellbeing using multi-dimensional frameworks, and incorporating information from subjective data.

References Abdelkrim, A. and Duclos, J.-Y. (2007). ‘DASP: Distributive Analysis Stata Package’. PEP, World Bank, UNDP and Université Laval. Online resource available at: http://dasp.ecn.ulaval.ca/ (accessed: 22 September 2010).

Ackland, R., Dowrick, S. and Freyens, B. (2008). ‘Measuring global poverty: Why PPPs matter’. Australian National University Department of Economics, unpublished manuscript. Online resource available at: http://voson.anu.edu.au/papers/GlobalPoverty_17Sep2008.pdf (accessed: 22 September 2010).

Ahluwalia, M. S., Cartner, N. G. and Chenery, H. B. (1979). ‘Growth and poverty in developing countries’. Journal of Development Economics 6, 299–341.

Alkire, S. and Santos, M. E. (2010). ‘Acute multi-dimensional poverty: A new index for developing countries’. UNDP Human Development Research Paper No. 11. New York: United Nations Development Programme.

Anand, S. and Segal, P. (2008). ‘What do we know about global income inequality?’ Journal of Economic Literature 46, 57-94.

Babones, S. J. (2003). ‘One world or two? A snapshot of the global income distribution’. Paper presented at the American Sociological Association 98th Annual Meeting (16−19 August), Atlanta, Georgia, USA.

Bandourian, R., McDonald, J. B. and Turley, R. S. (2003). ‘A comparison of parametric models of income distributions across countries and over time’. Revista Estadistica 55, 164-165.

Betson, D. M. (1996). ‘Is everything relative? The role of equivalence scales in poverty measurement’. University of Notre Dame Department of Economics, unpublished manuscript. Online resource available at: http://www.nd.edu/~dbetson/research/documents/EverythingRelative.pdf (accessed: 22 September 2010).

Betson, D. M. (2004). ‘Poverty equivalence scales: Adjustment for demographic differences across families’. Paper presented at the 2004 National Academy of Science conference and University of Notre Dame Department of Economics, unpublished manuscript. Online resource available at: http://www.nd.edu/~dbetson/research/documents/EquivalenceScales.pdf (accessed: 22 September 2010).

Bhalla, S. (2002). ‘Imagine there’s no country: Poverty, inequality and growth in the era of globalization’. Population Studies 55, 263–279.

Bongaarts, J. (2001). ‘Household size and composition in the developing world in the 1990s’. Population Studies 55, 263–279.

Bourguignon, F. and Morrison, C. (2002). ‘Inequality among world citizens: 1820-1992’. American Economic Review 92, 727–744.

Buhmann, B., Rainwater, L., Schmaus, G. and Smeeding, T. M. (1988). ‘Equivalence scales, wellbeing, inequality and poverty: Sensitivity estimates across ten countries using the Luxembourg Income Study (LIS)’. Review of Income and Wealth 34(2), 115-142.

Burniaux, J.-M., Dang, T.-T., Fore, D., Förster, M. F., Mira d’Ercole, M. and Oxley, H. (1998). ‘Income distribution and poverty in selected OECD countries’, OECD Economics Department Working Paper No. 189. Paris, France: Organization for Economic Co-operation and Development.

Chamon, M. D. and Prasad, E. S. (2010). ‘Why are savings rates of urban households in China rising?’ American Economic Journal: Macroeconomics 2, 93-130.

Chotikapanich, D., Griffiths, W. E. and Rao, D. S. P. (2007). ‘Estimating and combining national income distributions using limited data’, Journal of Business and Economics Statistics 25, 97-109.

Chotikapanich, D., Rao, D. S. P. and Tang, K. K. (2007). ‘Estimating income inequality in China using grouped data and the Generalized Beta distribution’, Review of Income and Wealth 53, 127- 147.

Chen, S., Datt, G. and Ravallion, M. (1994). ‘Is poverty increasing in the developing world?’ Review of Income and Wealth 40, 359-76.

Chen, S. and Ravallion, M. (2001). ‘How did the world’s poorest fare in the 1990s?’ Review of Income and Wealth 47, 283-300.

Chen, S. and Ravallion, M. (2004). ‘How have the world’s poorest fared since the early 1980s?’ World Bank Research Observer 19, 141-169.

Chen, S. and Ravallion, M. (2007). ‘Absolute poverty measures for the developing world, 1981– 2004’. World Bank Policy Research Working Paper Series No. 4211. Washington, DC: The World Bank Group.

Chen, S. and Ravallion, M. (2008). ‘The developing world is poorer than we thought, but no less successful in the fight against poverty’. Quarterly Journal of Economics, forthcoming.

Datt, G. (1998). ‘Computational tools for poverty measurement and analysis’. IFPRI Food Consumption and Nutrition Division Discussion Paper No. 50. Washington, DC: International Food Policy Research Institute.

Deaton, A. (2001). ‘Counting the world’s poor: Problems and possible solutions’. World Bank Research Observer 16, 125-147.

Deaton, A. (2003). ‘Household surveys, consumption, and the measurement of poverty’. Economic Systems Research 15, 165-159.

Deaton, A. (2005). ‘Measuring poverty in a growing world (or Measuring growth in a poor world)’. Review of Economics and Statistics 87, 1-19.

Deaton, A. (2008) ‘Income, health and well-being around the world: Evidence from the Gallup World Poll’. Journal of Economic Perspectives 22, 53-72.

Deaton, A. (2010). ‘Price indexes, inequality, and the measurement of global poverty’. Princeton, NJ: Princeton University, Department of Economics, unpublished manuscript.

Deaton, A. and Dupriez, O. (2008). ‘Poverty PPPs around the world: An update and progress report’. World Bank Development Data Group. Washington, DC: The World Bank Group.

Deaton A. and Grosh, M. (2000). ‘Consumption’. In M. Groshand P. Glewwe (eds.), Designing Household Questionnaires for Developing Countries: Lessons from Fifteen Years of the Living Standard Measurement Study. Washington, DC: The Word Bank Group.

Deaton, A. and Kozel, V. (2005). ‘Data and dogma: The great Indian poverty debate’. World Bank Research Observer 20, 177-199.

Deininger, K. and Squire, L. (1996). ‘A new data set measuring income inequality’. World Bank Economic Review 10, 565-591.

Dhongde, S. (2007). ‘Measuring the impact of growth and income distribution on poverty in India’. Journal of Income Distribution 16, 25-48.

Dollar, D. and Kraay, A. (2002). ‘Growth is good for the poor’. Journal of Economic Growth 7, 195- 225.

Dowrick, S. and Akmal, H. (2005). ‘Contradictory trends in global income inequality: A tale of two biases’. Review of Income and Wealth 51, 201-229.

Elekdag, S. and Lall, S. (2008). ‘Global growth estimates trimmed after PPP revisions’. IMF Survey Magazine. Washington, DC: International Monetary Fund. Online resource available at http://www.imf.org/external/pubs/ft/survey/so/2008/RES018A.htm (accessed: 22 September 2010).

Ferreira, F. and Ravallion, M. (2008a). ‘Global poverty and inequality: A review of the evidence’. World Bank Policy Research Working Paper Series No. 4623. Washington, DC: The World Bank Group.

Ferreira, F. and Ravallion, M. (2008b). ‘Poverty and inequality: The global context’. In W. Salverda, B. Nolan and T. Smeeding (eds.), Oxford Handbook of Economic Inequality. Oxford: Oxford University Press.

Gasparini, L. C. and Gluzmann, P. A. (2009). ‘Estimating income poverty and inequality from the Gallup World Poll: The case of Latin America and the Caribbean’. ECINEQ Working Paper No. 15. Palma de Mallorca, Spain: Society for the Study for Economic Inequality.

Haughton, J. and Khandker, S. R. (2009). Handbook on Poverty and Inequality. International Bank for Reconstruction and Development/ World Bank. Washington, DC: The World Bank Group.

Heston, A., Summers, R. and Atten, B. (2009). Penn World Table Version 6.3, Center for International Comparisons of Production, Income and Prices at the University of Pennsylvania. Online resource available at: http://pwt.econ.upenn.edu/php_site/pwt_index.php (accessed: 22 September, 2010).

Himanshu (2009). ‘New global poverty estimates: What do these mean?’ JNU Centre for Study of Regional Development, unpublished manuscript. New Delhi, India: Jawaharlal Nehru University.

Jann, B. (2005). ‘KDENS: Stata module for univariate kernel density estimation’. Online resource available at http://ideas.repec.org/c/boc/bocode/s456410.html (accessed: 22 September 2010).

Kakwani, N. (1980). ‘On a class of poverty measures’. Econometrica 44, 137–148.

Kakwani, N. and Son, H. (2006). ‘New global poverty counts’. UNDP International Poverty Center Working Paper No. 29. Brasilia, Brazil: United Nations Development Program.

Kapan, T. (2009). ‘Patriarchal households are unitary: New evidence’. Columbia University Department of Economics, unpublished manuscript. New York: Columbia University.

Klasen, S. (2009). ‘Levels and trends in absolute poverty in the world: What we know and what we don’t’. Universität Göttingen Courant Research Center Discussion Paper No. 11. Göttingen, Germany: Georg-August-Universität Göttingen.

Lise, J. and Seitz, S. (2008). ‘Consumption inequality and intra-household allocations’. Review of Economic Studies, forthcoming.

Milanovic, B. (2002). ‘True World Income Distribution, 1988 and 1993: First calculation based on household surveys alone’. The Economic Journal 112, 51-92.

Milanovic, B. (2005). Worlds Apart: Measuring International and Global Inequality. Princeton, NJ: Princeton University Press.

Milanovic, B. (2009). ‘Global inequality recalculated: The effect of the new 2005 PPP estimates on global inequality’. Journal of Economic Inequality, forthcoming. 20

Minoiu, C. and Reddy, S. (2008). ‘Kernel density estimation on grouped data: The case of poverty assessment’. IMF Working Paper No. 08/183. Washington, DC: International Monetary Fund.

Minoiu, C. and Reddy, S. (2009). ‘Estimating poverty and inequality from grouped data: How well do parametric methods perform?’ Journal of Income Distribution 18, 160-178.

Mistiaen, J. and Ravallion, M. (2003). ‘Survey compliance and the distribution of income’. World Bank Policy Research Working Paper No. 295. Washington, DC: The World Bank Group.

Nielsen, L. (2009). ‘Global relative poverty’. IMF Working Paper No. 09/93. Washington, DC: International Monetary Fund.

Paukert, F. (1973). ‘Income distribution at different levels of development: A survey of evidence’. International Labour Review 108, 97-125.

Pinkovskiy, M. and Sala-i-Martin, X. (2009). ‘Parametric distributions of the world distribution of income’. NBER Working Paper No. 15433. Cambridge, MA: The National Bureau for Economic Research.

Pritchett, L. (2006). ‘Who is not poor? Dreaming of a world truly free of poverty’. World Bank Research Observer 21, 1-23.

Quibria, M. G. (2005). ‘Measuring global poverty right’. World Economics 6, 111-121.

Ravallion, M. (2003). ‘Measuring aggregate welfare in developing countries: How well do national accounts and surveys agree?’ Review of Economics and Statistics 85, 645-652.

Ravallion, M. and Chen, S. (1997). ‘What can new survey data tell us about recent changes in distribution and poverty?’ World Bank Economic Review 11, 357-382.

Ravallion, M. and Chen, S. (2009). ‘Weakly relative poverty’. Review of Economics and Statistics, forthcoming.

Ravallion, M., Chen, S. and Sangraula, P. (2009). ‘Dollar a day revisited’. World Bank Economic Review, forthcoming.

Ravallion, M., Datt, G. and van de Walle, D. (1991). ‘Quantifying absolute poverty in the developing world’. Review of Income and Wealth 37, 345-361.

Reddy, S. and Minoiu, C. (2007). ‘Has world poverty really fallen?’ Review of Income and Wealth 53, 484-502.

Reddy, S. and Pogge, T. (2010). ‘How not to count the poor’. In Anand, S., Segal, P. and Stiglitz, J. (eds.), Debates on the Measurement of Poverty. Oxford: Oxford University Press. 21

Reddy, S., Visaria, S. and Assali, M. (2009). ‘Intercountry comparisons of poverty based on a capability approach’. In K. Basu and R. Kanbur (eds.), Arguments for a Better World: Essays in Honor of Amartya Sen. Oxford: Oxford University Press.

Sala-i-Martin, X. (2002a). ‘The world distribution of income (estimated from individual country distributions)’. NBER Working Paper No. 8933. Boston, MA: National Bureau of Economic Research.

Sala-i-Martin, X. (2002b). ‘The “disturbing” rise of world income inequality’. NBER Working Paper No. 8904. Boston, MA: National Bureau of Economic Research.

Sala-i-Martin, X. (2004). ‘The world distribution of income: Falling poverty and… convergence, period’. Columbia University Department of Economics, unpublished manuscript. Online resource available at: http://www.econ.upf.edu/~xsala/BMSS/XSiM%20WDI.pdf (accessed 20 October 2010).

Sala-i-Martin, X. (2006). ‘The world distribution of income: Falling poverty and… convergence, period’. Quarterly Journal of Economics 121, 351-397.

Sen, A. (1976). ‘Poverty: An ordinal approach to measurement’. Econometrica 44, 219-231.

Sen, A. (1993). ‘Capability and well-being’. In A. Sen and M. Nussbaum (eds.), The Quality of Life. Oxford: Clarendon Press.

Shorrocks, A. and Wan, G. (2008). ‘Ungrouping income distributions: Synthesizing samples for inequality and poverty analysis’. UNU-WIDER Research Paper No. 16. Helsinki, Finland: United Nations University World Institute for Development Economic Research.

Silverman, B. W. (1986). Density Estimation for Statistics and Data Analysis. Monographs on Statistics and Applied Probability 26. Boca Raton, FL; London; New York: Chapman & Hall/CRC.

Szekely, M., Lustig, N., Cumpa, M. and Media-Guerra, J. A. (2004). ‘Do we know how much poverty there is?’ Oxford Development Studies 4, 523−558.

Trinczek, M. (2007). ‘Income poverty in the European Union’. European Foundation for the Improvement of Living and Working Conditions Online resource available at http://www.eurofound.europa.eu/ewco/surveyreports/EU0703019D/EU0703019D.pdf (accessed: 22 September 2010).

UNU-WIDER. World Income Inequality Database (WIID) (2008). Helsinki, Finland: United Nations University World Institute for Development Economics Research. Online resource available at http://www.wider.unu.edu/research/Database/en_GB/database/ (accessed: 22 September 2010).

Villasenor, J. A. and Arnold, B. C. (1989). ‘Elliptical Lorenz curves’. Journal of Econometrics 40, 327-338.

Wand, M. P. and Jones, M. C. (1995). Kernel Smoothing. London: Chapman and Hall.

Wei, S.-J., and Zhang, X. (2009). ‘The competitive savings motive: Evidence from rising sex ratios and savings rates in China’. NBER Working Paper No. 15093. Cambridge, MA: The National Bureau of Economic Research.

World Bank (2000). The World Development Report (WDR). Washington, DC: The World Bank Group.

World Development Indicators (WDI) (2006). Washington, DC: The World Bank Group.

World Development Indicators (WDI) (2009). Washington, DC: The World Bank Group.

Wu, X. and Perloff, J. (2005). ‘China’s income distributions: 1985−2001’. Review of Economics and Statistics 87, 763-775.

Wu, X. and Perloff, J. (2007). ‘GMM estimation of a maximum entropy distribution with interval data’. Journal of Econometrics 138, 532-546.

Zhang, Y. and Wan, G. (2006). ‘Globalization and the urban poor in China’. UNU-WIDER Working Paper No. 42. Helsinki, Finland: United Nations University World Institute for Development Economics Research.

VI. Appendix

Table 1. Demographic trends and gender inequalities by region, 1970–2004

% change Demographic trends 1970 1980 1990 2004 1970-2004 Population ages < 14 (% of total) East Asia & Pacific 40.6 36.9 30.3 24.5 -39.7 Europe & Central Asia 29.3 26.3 26.4 20.2 -30.8 Latin America & Caribbean 42.5 39.8 36.4 30.4 -28.4 Middle East & North Africa 45.1 44.3 43.3 34.0 -24.6 South Asia 41.0 39.5 37.8 33.8 -17.6 Sub-Saharan Africa 44.8 45.4 45.6 43.7 -2.4 Population ages > 65 (% of total) East Asia & Pacific 4.1 4.5 5.1 6.8 66.2 Europe & Central Asia 7.8 9.4 9.0 11.6 48.9 Latin America & Caribbean 4.2 4.4 4.7 5.9 42.3 Middle East & North Africa 3.8 3.6 3.5 4.2 10.4 South Asia 3.6 3.9 4.1 4.8 34.4 Sub-Saharan Africa 2.9 2.9 2.9 3.1 4.7 Average household size Asia 5.14 Europe & Central Asia 3.60 Latin America & Caribbean 4.76 Middle East & North Africa 5.65 Sub-Saharan Africa 5.25 Source: World Development Indicators (2006) and authors’ calculations.

Table 2. Gender inequalities by region, 1990–2004

% change Gender inequalities 1990 2004 1990-2004

Female participation in the labour force East Asia & Pacific 44.10 43.84 -0.6 Europe & Central Asia 45.59 44.93 -1.4 Latin America & Caribbean 34.03 40.24 18.2 Middle East & North Africa 22.94 27.17 18.5 South Asia 30.64 29.33 -4.3 Sub-Saharan Africa 43.03 42.19 -1.9 Ratio of male to female primary completion rates East Asia & Pacific 1.02 1.02 0.0 Europe & Central Asia 0.99 1.01 1.5 Latin America & Caribbean 0.96 0.99 2.3 Middle East & North Africa 1.17 1.04 -11.3 South Asia 1.33 1.09 -17.8 Sub-Saharan Africa 1.17 1.18 1.0 Ratio of male to female literacy rates East Asia & Pacific 1.24 1.09 -12.2 Europe & Central Asia 1.04 1.01 -3.1 Latin America & Caribbean 1.03 1.02 -1.8 Middle East & North Africa 1.67 1.26 -24.6 South Asia 1.76 1.55 -11.8 Sub-Saharan Africa 1.50 1.30 -13.5 Source: World Development Indicators (2006), Bongaarts (2001) for average household size, and authors’ calculations.

Table 3. Chronology of global poverty studies % of developing # of countries # of surveys Poverty estimates are reported in: world population Grouped data and individual records Ahluwalia, Carter, and Chenery (1979) 361/ 1975 Ravallion, Datt, and van de Walle (1991) 222/ 1985 Chen, Datt, and Ravallion (1994) 40 1985, 1990 Ravallion and Chen (1997) 67 85% 109 1987, 1990, 93 Chen and Ravallion (2001) 83 88% 265 1987, 1990, 93, 96, 98 Chen and Ravallion (2004) 97 93% 454 1981, 84, 87, 1990, 93, 96, 99, 2001 Chen and Ravallion (2007) 100 93% 500 1981, 84, 87, 1990, 93, 96, 99, 2002, 04 Chen and Ravallion (2008) 115 90%3/ 675 1981, 84, 87, 1990, 93, 96, 99, 2002, 05 Grouped data Dowrick and Akmal (2005) 47 70% 1980, 1993 1850, 1970, 1890, 1910, 29, 59, 60, 70, Bourguignon and Morrisson (2002) 33 80, 92 Sala-i-Martin (2002a, 2002b) 97+284/ 90% 1970, 1980, 1990, 1998 Sala-i-Martin (2004) 111+284/ 93% 1970, 75, 1980, 85, 1990, 95, 2000 Sala-i-Martin (2006) 110+284/ 93% 1970-2000 Pinkovskiy and Sala-i-Martin (2009) 191 98% 1069 1970-2006 Individual records Milanovic (2002) 91, 119 86%, 91% 216 1988, 1993 1/ "Fairly reliable" data was only available for 25 of the 36 countries (Ahluwalia, Cartner, and Chenery, 1979, footnote. 1); for the remainder, the data was estimated using cross-country comparisons. 2/ Poverty is estimated for an additional 64 countries using an extrapolation model. 3/ The coverage varies between 74 percent for the MENA region; and 98 percent for SSA. 4/ Distributional information exists for 97 (or 111) countries, depending on the study. It is imputed for 28 countries.

Table 4. Data available for global analyses, 1960–2007 # surveys per DATABASE # countries # surveys coverage country World Income Inequality Database (WIID) 140 2211 15.8 1960-2006 Povcalnet 1/ 116 447 3.9 1980-2007 World Development Indicators (WDI) 90 162 1.8 1982-2005 Dollar and Kraay (2002) 79 211 2.7 1961-1999 Total 154 3031 19.7 1960-2007 of which : Low-income 39 369 9.5 Middle-income 78 1629 20.9 High-income 37 1027 27.8 of which: East Asia & Pacific 12 209 17.4 Europe & Central Asia 22 450 20.5 Latin America & Caribbean 24 757 31.5 Middle East & North Africa 9 90 10.0 OECD and high-income non-OECD 37 1027 27.8 South Asia 6 165 27.5 Sub-Saharan Africa 44 327 7.4 Source: Authors’ calculations. 1/ Povcalnet reports rural and urban distributional data for three countries: China, India and Indonesia.

Figure 1. Data availability (WIID and Povcalnet), 1960–2005

Country-surveys from WIID Country-surveys from Povcalnet 180 Number of countries (Full dataset) 160

140

120

100

60 Number ofcountry-surveys Number

0 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005

Source: Authors’ calculations.

Figure 2. Global poverty rates from different sources (%), 1981–2005 1993 PPPs 2005 PPPs 60 60 Chen and Ravallion (2008) 50 Bhalla (2002) 50 Sala-i-Martin (2006) 40 40

30 30

20 20 Chen and Ravallion (2008)

Pinkovskiy and Sala-i-Martin (2009) 10 10

0 0 1981 1987 1993 1999 2005 1981 1987 1993 1999 2005 Notes: The various estimates are not strictly comparable because of differences in methodologies (see text). Furthermore, Chen and Ravallion (2008) and Bhalla (2002) compute the $1/day poverty rates relative to the developing world population, whereas Sala-i-Martin uses the total world population as denominator. The figures are for the year listed or the closest available year.

Table 5. Methodological differences between recent studies Characteristics of the analysis Chen and Ravallion (2008) Pinkovskiy and Sala-i-Martin (2009) Scope of analysis Developing world World Number of countries 115 191 % of (developing) world population 90 97.9 Number of surveys 675 1069 Type of data Individual records, grouped data Grouped data Interpolation/extrapolation techniques Yes, to line up surveys with Yes, of Gini coefficients for reference years missing years Welfare indicator Consumption Income Source of data for welfare indicator HS; when HS unavailable, use NAS NAS with adjustment International poverty line $1.25/day (at 2005 PPP) $1/day to $10/day (at 2005 PPP) Estimation method Generalized Quadratic (or Beta) Log-normal assumption for the for the Lorenz curve density

Figure 3. Impact of survey vs. national accounts mean income/consumption on the global distribution

1995 2005 .5 .5 .4 .4 .3 .3 .2 .2 .1 .1 0 0

2 4 6 8 10 12 2 4 6 8 10 12

HS-CONS NAS-GDP HS-CONS NAS-GDP NAS-CONS NAS-CONS

Source: Authors’ calculations. Note: The world income distribution has been obtained by integrating national distributions estimated, assuming zero within-quintile inequality; and smoothed using a kernel density estimator with Gaussian kernel and optimal bandwidth (Silverman, 1986). The vertical line is placed at $1/day international poverty line.

Table 6. List of countries included in the sensitivity analysis (N=65) Albania Dominican Rep. Kenya Panama Argentina Ecuador Kyrgyz Republic Paraguay Armenia Egypt Latvia Peru Azerbaijan El Salvador Lithuania Philippines Bangladesh Estonia Madagascar Poland Belarus Ethiopia Malawi Romania Bolivia Georgia Malaysia Russian Federation Brazil Guinea Mali Senegal Bulgaria Honduras Mexico Slovenia Burkina Faso Hungary Moldova, Rep. Thailand Cambodia India Mongolia Turkey Central African Republic Indonesia Nepal Uganda Chile Iran Nicaragua Ukraine China Jamaica Niger Uruguay Colombia Jordan Nigeria Venezuela, RB Costa Rica Kazakhstan Pakistan Vietnam Zambia

Table 7. Summary statistics for survey and national accounts mean income/consumption Mean St. Dev. Min Max Mean St. Dev. Min Max Unweighted 1995 2005 HS-CONS 1,922 1,350 287 6,668 2,148 1,512 409 8,241 NAS-CONS 3,095 1,879 625 8,205 4,104 2,611 624 11,714 NAS-GDP 4,957 3,431 791 13,436 6,669 4,927 834 22,004

Population-weighted HS-CONS 1,222 998 287 6,668 1,331 1,065 409 8,241 NAS-CONS 2,256 1,421 625 8,205 3,051 1,733 624 11,714 NAS-GDP 3,877 2,377 791 13,436 5,614 3,064 834 22,004 Source: Povcalnet for survey-based consumption (HS-CONS) and Penn World Tables (PWT) Mark 6.3 for national accounts-based consumption (NAS-CONS) and income (NAS-GDP). All figures expressed in 2005 international US$ (chain-weighted series in the case of PWT data).

Table 8. Sensitivity of global poverty estimates to data source for mean income or consumption (survey vs. national accounts)

1995 2005 Reduction over 1995-2005 HS-CONS NAS-CONS NAS-GDP HS-CONS NAS-CONS NAS-GDP HS-CONS NAS-CONS NAS-GDP Headcount ratio (%) (% ) $1.00/day 29.0 5.9 1.4 24.3 1.7 0.9 -16 -32 -72 $1.25/day 38.6 10.7 2.7 33.7 2.9 1.5 -13 -44 -73 $1.45/day 45.1 14.8 4.2 40.2 5.0 2.0 -11 -53 -66 $2.00/day 58.5 25.8 9.6 54.2 13.5 3.7 -7 -62 -47 $2.50/day 66.6 35.1 15.6 62.8 21.4 5.5 -6 -65 -39

Absolute headcount (millions) (millions) $1.00/day 1,219 250 58 1,140 78 44 -6 -24 -69 $1.25/day 1,621 452 112 1,579 136 70 -3 -37 -70 $1.45/day 1,893 620 177 1,887 234 93 0 -48 -62 $2.00/day 2,458 1,082 405 2,540 635 174 3 -57 -41 $2.50/day 2,798 1,476 654 2,945 1,002 259 5 -60 -32 Source: Authors’ calculations using the Generalised Quadratic parameterisation of the Lorenz curve (Vilasenor and Arnold, 1989; Datt, 1998). Estimates obtained using the Stata package DASP Version 2.1 (Abdelkrim and Duclos, 2007).

Table 9. Sensitivity of global poverty estimates to the estimation method (selected poverty lines) 1995 2005 Reduction over 1995-2005

$1.00/day $1.45/day $2.50/day $1.00/day $1.45/day $2.50/day $1.00/day $1.45/day $2.50/day Headcount ratio (%) (%) Parametric methods GQ 5.9 14.8 35.1 1.7 5.0 21.4 -72 -66 -39 Beta 4.2 11.0 29.9 1.3 3.3 17.6 -70 -70 -41 Beta* 4.4 11.1 30.3 1.3 3.7 17.0 -70 -67 -44 Log-normal 5.3 11.8 29.7 2.3 5.7 17.6 -57 -52 -41 Log-normal* 4.3 11.0 30.3 1.6 4.2 16.9 -62 -62 -44 Singh-Maddala 5.5 12.3 34.0 2.3 5.6 20.6 -58 -55 -39 Nonparametric methods Kernel Silverman 5.4 12.5 31.4 1.6 5.2 19.0 -69 -58 -40 Kernel Normalscale 6.0 13.1 31.7 2.0 5.7 19.2 -67 -56 -39 Kernel Oversmooth 8.9 16.6 32.9 4.1 9.2 22.7 -55 -45 -31 Kernel DPI-1 7.6 14.9 32.2 2.8 7.2 20.6 -64 -52 -36 Kernel DPI-2 8.1 15.6 32.2 3.2 7.7 21.5 -60 -51 -33 Kernel DPI-3 8.6 15.7 32.2 3.4 7.9 21.6 -61 -49 -33 Absolute headcount (millions) (millions) Parametric methods GQ 250 620 1,476 78 234 1,002 -171 -385 -474 Beta 176 461 1,257 59 156 827 -116 -304 -430 Beta* 183 468 1,273 62 174 796 -121 -294 -477 Log-normal 225 497 1,247 108 268 827 -116 -229 -420 Log-normal* 179 462 1,273 76 197 793 -103 -265 -480 Singh-Maddala 232 518 1,430 109 263 966 -123 -256 -464 Nonparametric methods Kernel Silverman 226 527 1,321 77 244 891 -149 -282 -430 Kernel Normalscale 254 551 1,332 95 269 901 -159 -282 -431 Kernel Oversmooth 375 697 1,381 190 429 1,066 -185 -268 -315 Kernel DPI-1 318 625 1,351 129 335 967 -189 -290 -384 Kernel DPI-2 340 655 1,355 150 361 1,008 -190 -294 -347 Kernel DPI-3 362 658 1,354 157 372 1,012 -205 -286 -342 Source: Authors’ calculations for selected poverty lines. Estimates obtained using the Stata package DASP Version 2.1 (Abdelkrim and Duclos, 2007) for parametric methods; and the KDENS Stata command (Jann, 2005) for the non-parametric (kernel) approach. The welfare aggregate is NAS-consumption. * labels estimates obtained using the Shorrocks and Wan (2008) iterative correction procedure, which ensures that the sample moments of the estimated distribution are the same as those of the raw data. Six estimates are presented for the kernel density estimator (applied to quintile means), each corresponding to a different ‘optimal’ bandwidth (in italics). These include the Silverman (1986) bandwidth, the normalscale bandwidth – a variant of Silverman (1986) which assumes normality of the log- income distribution, the oversmooth bandwidth (a good starting point for bandwidth fine-tuning) and three direct-plug in (DPI) bandwidths that iteratively estimate the density (Wand and Jones, 1995).

Table 10. Variation in global poverty estimates due to estimation method $1.00/day $1.45/day $2.50/day $1.00/day $1.45/day $2.50/day $1.00/day $1.45/day $2.50/day Headcount ratio Ratio between maximum and minimum 2.1 1.5 1.2 3.2 2.7 1.3 0.8 0.6 0.7 estimate Difference between max and min estimate 5 6 5 3 6 6 17 25 13 (percentage points) Absolute headcount Ratio between maximum and minimum 2.1 1.5 1.2 3.2 2.7 1.3 0.5 0.6 0.7 estimate Difference between max and min estimate 200 237 229 131 273 272 102 157 165 (millions) Source: Authors’ calculations.

The Brooks World Poverty Institute (BWPI) creates and

shares knowledge to help end global poverty. Executive Director

Professor David Hulme BWPI is multidisciplinary, researching poverty in both Research Director the rich and poor worlds. Professor Armando Barrientos

Our aim is to better understand why people are poor,

what keeps them trapped in poverty and how they can

be helped - drawing upon the very best international

Contact: practice in research and policy making.

Brooks World Poverty Institute The University of Manchester The Brooks World Poverty Institute is chaired by Nobel Humanities Bridgeford Street Building Laureate, Professor Joseph E. Stiglitz. Oxford Road Manchester M13 9PL

United Kingdom

Email: [email protected] www.manchester.ac.uk/bwpi www.manchester.ac.uk/bwpi